All notable changes to this project will be documented in this file.
The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.
In particular:
- Any updates to pre-existing files shall entail a major version bump.
- The inclusion of a new language, without changes to other pre-existing data, shall only entail a minor version bump.
It follows from the above that benchmark scores across different releases of FLORES+ with the same major version number are comparable.
- Updated
nob_Latn
after additional quality assessment.
- Relabeled macrolanguage codes to the correct individual language codes:
est_Latn
toekk_Latn
,grn_Latn
togug_Latn
,kon_Latn
toktu_Latn
. - The
ajp
code has been deprecated. Relabeledajp_Arab
toapc_Arab_sout3123
andapc_Arab
toapc_Arab_nort3139
. - Relabeled Asante and Akuapem data:
aka_Latn
totwi_Latn_asan1239
andtwi_Latn
totwi_Latn_akua1239
.
- Added
nqo_Nkoo
from https://github.com/common-parallel-corpora/common-parallel-corpora. - Added the
dev
splits forbrx_Deva
,dgo_Deva
,mni_Mtei
,snd_Deva
,gom_Deva
from https://github.com/ai4bharat/IndicTrans2. - Added the
dev
split formhr_Cyrl
. Many thanks to Andrey Chemyshev @fu-lab. - Added the
dev
split forchv_Cyrl
. Many thanks to @AlAntonov.
- Relabeled
zho_Hans
tocmn_Hans
after confirming the data is in Standard Beijing Mandarin. - Relabeled
zho_Hant
tocmn_Hant
after confirming the data is in Taiwanese Mandarin. - Relabeled
tgl_Latn
tofil_Latn
after confirming the data is in Filipino. - Relabeled
tzm_Tfng
tozgh_Tfng
after additional quality assessment revealed the data was in Standard Moroccan Tamazight. Many thanks to @MedAymenF for pointing out the issues with the original FLORES-200 data. - Updated
lij_Latn
after additional quality assessment. Data has undergone minor spelling and syntactic fixes. - Updated
ckb_Arab
to replace usage of the non-standardك
character and improve the translation quality. Many thanks to @Sarchia for pointing out the issues with the original FLORES-200 data. - Updated
yue_Hant
to better conform to Honk Kong Cantonese. Thanks to the users who reported issues in facebookresearch/flores#61. - Updated several datasets due to a small number of translations (typically 1-2 sentences per dataset) having been overwritten by other sentences. The issue was reported in facebookresearch/flores#62 and facebookresearch/flores#67 – many thanks to @sotwi and @kargaranamir! Affected dev sets:
ary_Arab
,azb_Arab
,ban_Latn
,bod_Tibt
,bos_Latn
,bug_Latn
,crh_Latn
,dik_Latn
,dyu_Latn
,dzo_Tibt
,fao_Latn
,hat_Latn
,jav_Latn
,kam_Latn
,kas_Deva
,kaz_Cyrl
,kbp_Latn
,lim_Latn
,lin_Latn
,lit_Latn
,lus_Latn
,npi_Deva
,run_Latn
,san_Deva
,sat_Olck
,spa_Latn
,ssw_Latn
,sun_Latn
,szl_Latn
,taq_Tfng
,urd_Arab
,ydd_Hebr
. Affected devtest sets:ary_Arab
,bam_Latn
,bod_Tibt
,dyu_Latn
,gla_Latn
,grn_Latn
,hat_Latn
,hne_Deva
,kam_Latn
,kaz_Cyrl
,kik_Latn
,lin_Latn
,lit_Latn
,lua_Latn
,min_Arab
,min_Latn
,npi_Deva
,run_Latn
,san_Deva
,smo_Latn
,spa_Latn
,szl_Latn
,taq_Tfng
,tgk_Cyrl
,uig_Arab
,urd_Arab
,ydd_Hebr
.
Initial release. This is the exact same data that was released under the name FLORES-200
by NLLB Team et al. (2022). It can be downloaded from https://tinyurl.com/flores200dataset.